perf: also use the main domain, along with n-1 other ones

we always keep a thread alive on the main domain as a worker for new
tasks, but other domains can still come and go to manage resources
properly in case a pool is started and used only for a short while.

src/d_pool_.ml

@@ -18,9 +18,7 @@ type worker_state = {
     including a work queue and a thread refcount; and the domain itself,
     if any, in a separate option because it might outlive its own state. *)
 let domains_ : (worker_state option * Domain_.t option) Lock.t array =
-  (* number of domains we spawn. Note that we spawn n-1 domains
-      because there already is the main domain running. *)
-  let n = max 1 (Domain_.recommended_number () - 1) in
+  let n = max 1 (Domain_.recommended_number ()) in
   Array.init n (fun _ -> Lock.create (None, None))
 
 (** main work loop for a domain worker.
@@ -84,6 +82,14 @@ let work_ idx (st : worker_state) : unit =
   done;
   ()
 
+(* special case for main domain: we start a worker immediately *)
+let () =
+  assert (Domain_.is_main_domain ());
+  let w = { th_count = Atomic_.make 1; q = Bb_queue.create () } in
+  (* thread that stays alive *)
+  ignore (Thread.create (fun () -> work_ 0 w) () : Thread.t);
+  domains_.(0) <- Lock.create (Some w, None)
+
 let[@inline] n_domains () : int = Array.length domains_
 
 let run_on (i : int) (f : unit -> unit) : unit =

src/domain_.ml

@@ -9,6 +9,7 @@ let get_id (self : t) : int = (Domain.get_id self :> int)
 let spawn : _ -> t = Domain.spawn
 let relax = Domain.cpu_relax
 let join = Domain.join
+let is_main_domain = Domain.is_main_domain
 
 [@@@ocaml.alert "+unstable"]
 [@@@else_]
@@ -21,5 +22,6 @@ let get_id (self : t) : int = Thread.id self
 let spawn f : t = Thread.create f ()
 let relax () = Thread.yield ()
 let join = Thread.join
+let is_main_domain () = true
 
 [@@@endif]